What Interconnects is
Interconnects is a newsletter written by Nathan Lambert that covers the open-weight AI ecosystem — the world of AI models whose weights are publicly released, as opposed to the closed, proprietary systems from the biggest labs. If you're trying to understand which open models matter, how they compare to closed ones, and what the broader industry dynamics mean, Interconnects is one of the most cited places practitioners turn.
Think of it as a well-informed colleague who reads everything, visits AI labs in person, and isn't afraid to say when the conventional wisdom is wrong.
Why it matters
The AI landscape moves fast, and most coverage is either press releases from labs or highly technical papers. Interconnects sits in between: it synthesizes what's happening, explains what it means, and takes positions. That combination — informed analysis with a point of view — makes it useful for anyone trying to make sense of a field that produces major model releases at a "flagship-after-flagship cadence," as Lambert himself described one recent stretch.
It's also one of the few outlets that takes open-weight models seriously as a distinct category, rather than treating them as simply "the cheaper version" of closed models.
What it covers
Open-weight model roundups. The recurring "Latest open artifacts" series (numbered issues) surveys each new wave of releases in one place. Recent editions have covered Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1, Qwen 3.5, Nvidia's Nemotron Super, and models from Indian lab Sarvam — tracking not just what launched but what it means for the ecosystem.
The open-vs-closed debate. A central theme is how open and closed models relate to each other. Lambert has argued that they're on different exponentials — distinct capability and value curves — rather than a single race where open models are simply behind. He's also examined the structural dynamics of "perpetual catch-up," including distillation (training on outputs from frontier models) as a mechanism, and pushed back on what he calls alarmist framing around "distillation attacks."
Post-training and technique deep-dives. Interconnects covers the less-visible work that happens after a model is trained — the fine-tuning, preference optimization, and reinforcement learning that shape how a model actually behaves. A practitioner interview series (also numbered) brings in guests like Finbarr Timbers to discuss frontier post-training recipes firsthand.
AI governance and policy. Lambert engages directly with policy debates. He co-authored an op-ed with Kevin Xu arguing against banning open-source AI, and published a piece declaring that AI governance has entered an "AGI era" — framing it as a one-way transition the field wasn't ready for.
On-the-ground reporting. Interconnects occasionally goes beyond commentary: a piece offering firsthand notes from visits to leading AI labs in China provided rare insider perspective on research culture and strategic direction there.
How it takes positions
Interconnects doesn't just describe — it argues. Some recurring stances visible in the events:
- Benchmark scores are not the primary driver of open model success or adoption.
- AI self-improvement is real but "lossy," making fast-takeoff scenarios unlikely.
- Open-weight fearmongering is often misguided, including narratives around models like Claude Mythos.
- The open-weights ecosystem benefits from compounding effects when many participants contribute — a dynamic Lambert has argued China's open-first ecosystem exploits particularly well.
Recent focus
Through mid-2026, Interconnects has been tracking a dense period of open-weight releases from Chinese labs (GLM, Qwen, MiniMax, Kimi), analyzing what makes models like Gemma 4 succeed beyond their benchmark numbers, and watching second-tier labs like Zyphra, Cohere, and Poolside enter the open ecosystem. It has also been grappling with what frontier-level AI means for governance — a question Lambert frames as newly urgent.
Who it's for
Interconnects is written for people who are technically literate but don't need every paper explained from scratch — ML practitioners, researchers, product teams, and policy-adjacent readers who want a clear-eyed take on where open-weight AI is going and why it matters.
